Python core.VGDLParser类代码示例

def test6():
    """ Now with memory!"""
    from numpy import ndarray
    from examples.gridphysics.mazes import polarmaze_game
    from pybrain.optimization import SNES
    g = VGDLParser().parseGame(polarmaze_game)
    g.buildLevel(cheese_maze)
    game_env = GameEnvironment(g)
    net = buildNet(game_env.outdim, 10, 4, temperature=0.1, recurrent=True)
    
    algo = SNES(lambda x: someEpisodes(game_env, x, avgOver=6, maxSteps=30, exploretoo=False), net, verbose=True, desiredEvaluation=0.85)
    print algo.batchSize
    rows, cols = 2,3
    episodesPerStep = 5
    for i in range(rows*cols):
        pylab.subplot(rows, cols, i+1)
        algo.learn(episodesPerStep)
        if isinstance(algo.bestEvaluable, ndarray):
            net._setParameters(algo.bestEvaluable)
        else:
            net = algo.bestEvaluable
        plotBackground(game_env)    
        plotTrajectories(game_env, net)
        pylab.title(str((i+1)*episodesPerStep))
        if algo.desiredEvaluation <= algo.bestEvaluation:
            break
        print
    pylab.show()

def plotLSPIValues(gametype, layout, discountFactor=0.9, useTD=False, showValue=False):
    # build the game
    g = VGDLParser().parseGame(gametype)
    g.buildLevel(layout)
    
    # transform into an MDP and the mapping to observations
    C = MDPconverter(g)
    Ts, R, fMap = C.convert()    
    
    # find the the best least-squares approximation to the policy,
    # given only observations, not the state information
    if useTD:
        # state-based
        _, Tlspi = LSTD_PI_policy(fMap, Ts, R, discountFactor=discountFactor)
    else:
        # state-action-based
        _, Tlspi = LSPI_policy(fMap, Ts, R, discountFactor=discountFactor)
    
    # evaluate the policy
    Vlspi = trueValues(Tlspi, R, discountFactor=discountFactor)
        
    # plot those values    
    featurePlot((g.width, g.height), C.states, Vlspi)
    
    if showValue:
        # expected discounted reward at initial state
        Vinit = Vlspi[C.initIndex()]
        pylab.xlabel("V0=%.4f"%Vinit)

def testAugmented():
    from vgdl.core import VGDLParser
    from pybrain.rl.experiments.episodic import EpisodicExperiment
    from vgdl.mdpmap import MDPconverter
    from vgdl.agents import PolicyDrivenAgent
    
    
    zelda_level2 = """
wwwwwwwwwwwww
wA wwk1ww   w
ww  ww    1 w
ww     wwww+w
wwwww1ww  www
wwwww  0  Gww
wwwwwwwwwwwww
"""

    
    from examples.gridphysics.mazes.rigidzelda import rigidzelda_game
    g = VGDLParser().parseGame(rigidzelda_game)
    g.buildLevel(zelda_level2)
    env = GameEnvironment(g, visualize=False,
                          recordingEnabled=True, actionDelay=150)
    C = MDPconverter(g, env=env, verbose=True)
    Ts, R, _ = C.convert()
    print C.states
    print Ts[0]
    print R
    env.reset()
    agent = PolicyDrivenAgent.buildOptimal(env)
    env.visualize = True
    env.reset()
    task = GameTask(env)    
    exper = EpisodicExperiment(task, agent)
    exper.doEpisodes(1)

def test4():
    from numpy import ndarray
    from examples.gridphysics.mazes import polarmaze_game
    from pybrain.optimization import SNES, WeightGuessing
    g = VGDLParser().parseGame(polarmaze_game)
    g.buildLevel(labyrinth2)
    game_env = GameEnvironment(g)
    net = buildNet(game_env.outdim, 5, 4, temperature=0.1, recurrent=False)
    
    algo = SNES(lambda x: someEpisodes(game_env, x, avgOver=3), net, verbose=True, desiredEvaluation=0.75)
    #algo = WeightGuessing(lambda x: someEpisodes(game_env, x), net, verbose=True, desiredEvaluation=0.78)
    rows, cols = 2,2
    episodesPerStep = 4
    for i in range(rows*cols):
        pylab.subplot(rows, cols, i+1)
        algo.learn(episodesPerStep)
        if isinstance(algo.bestEvaluable, ndarray):
            net._setParameters(algo.bestEvaluable)
        else:
            net = algo.bestEvaluable
        plotBackground(game_env)    
        plotTrajectories(game_env, net)
        pylab.title(str((i+1)*episodesPerStep))
        if algo.desiredEvaluation <= algo.bestEvaluation:
            break
        print
    pylab.show()

def testRolloutVideo(actions=[0, 0, 2, 2, 0, 3] * 2):        
    from examples.gridphysics.mazes import polarmaze_game, maze_level_1
    from vgdl.core import VGDLParser
    from vgdl.tools import makeGifVideo
    game_str, map_str = polarmaze_game, maze_level_1
    g = VGDLParser().parseGame(game_str)
    g.buildLevel(map_str)
    makeGifVideo(GameEnvironment(g, visualize=True), actions)

def testRollout(actions=[0, 0, 2, 2, 0, 3] * 20):        
    from examples.gridphysics.mazes import polarmaze_game, maze_level_1
    from vgdl.core import VGDLParser
    game_str, map_str = polarmaze_game, maze_level_1
    g = VGDLParser().parseGame(game_str)
    g.buildLevel(map_str)    
    env = GameEnvironment(g, visualize=True, actionDelay=100)
    env.rollOut(actions)

def _createVGDLGame( gameSpec, levelSpec ):
    import uuid
    from vgdl.core import VGDLParser
    # parse, run and play.
    game = VGDLParser().parseGame(gameSpec)
    game.buildLevel(levelSpec)
    game.uiud = uuid.uuid4()
    return game

def test2():
    from examples.gridphysics.mazes import polarmaze_game, maze_level_1
    from vgdl.core import VGDLParser
    game_str, map_str = polarmaze_game, maze_level_1
    g = VGDLParser().parseGame(game_str)
    g.buildLevel(map_str)    
    actions = [1, 0, 0, 3, 0, 2, 0, 2, 0, 0, 0]
    env = GameEnvironment(g, visualize=True, actionDelay=100)
    env.rollOut(actions)
    env.reset()
    senv = SubjectiveGame(g, actionDelay=1500)
    senv.rollOut(actions)

def testStochMaze():
    from vgdl.core import VGDLParser
    from examples.gridphysics.mazes.stochastic import stoch_game, stoch_level
    g = VGDLParser().parseGame(stoch_game)
    g.buildLevel(stoch_level)
    C = MDPconverter(g, verbose=True)
    Ts, R, fMap = C.convert()
    print C.states
    print R
    for T in Ts:
        print T
    print fMap

def testMaze():
    from vgdl.core import VGDLParser
    from examples.gridphysics.mazes import polarmaze_game, maze_level_1
    game_str, map_str = polarmaze_game, maze_level_1
    g = VGDLParser().parseGame(game_str)
    g.buildLevel(map_str)
    C = MDPconverter(g, verbose=True)
    Ts, R, fMap = C.convert()
    print C.states
    print R
    for T in Ts:
        print T
    print fMap

def test4():
    """ Same thing, but animated. """
    from examples.gridphysics.mazes.windy import windy_stoch_game, windy_level
    from pybrain.rl.experiments.episodic import EpisodicExperiment
    from vgdl.interfaces import GameEnvironment, GameTask
    from vgdl.agents import PolicyDrivenAgent 
    g = VGDLParser().parseGame(windy_stoch_game)
    g.buildLevel(windy_level)
    env = GameEnvironment(g, visualize=True, actionDelay=100)
    task = GameTask(env)
    agent = PolicyDrivenAgent.buildOptimal(env)
    exper = EpisodicExperiment(task, agent)
    res = exper.doEpisodes(5)
    print res

def testPolicyAgent():
    from pybrain.rl.experiments.episodic import EpisodicExperiment
    from vgdl.core import VGDLParser
    from examples.gridphysics.mazes import polarmaze_game, maze_level_2
    from vgdl.agents import PolicyDrivenAgent
    game_str, map_str = polarmaze_game, maze_level_2
    g = VGDLParser().parseGame(game_str)
    g.buildLevel(map_str)
    
    env = GameEnvironment(g, visualize=False, actionDelay=100)
    task = GameTask(env)
    agent = PolicyDrivenAgent.buildOptimal(env)
    env.visualize = True
    env.reset()
    exper = EpisodicExperiment(task, agent)
    res = exper.doEpisodes(2)
    print res

def test1():
    from examples.gridphysics.mazes import polarmaze_game, maze_level_1    
    
    game_str, map_str = polarmaze_game, maze_level_1
    g = VGDLParser().parseGame(game_str)
    g.buildLevel(map_str)
    
    game_env = GameEnvironment(g)
    print 'number of observations:', game_env.outdim
    
    net = buildNet(game_env.outdim, 2, 2)
    for i in range(200):
        net.randomize()
        net.reset()
        print someEpisodes(game_env, net),
        if i% 20 == 19:
            print

def test3():
    from examples.gridphysics.mazes import polarmaze_game
    from examples.gridphysics.mazes.simple import maze_level_1b
    from vgdl.core import VGDLParser
    from pybrain.rl.experiments.episodic import EpisodicExperiment
    from vgdl.interfaces import GameTask
    from vgdl.agents import InteractiveAgent, UserTiredException
    game_str, map_str = polarmaze_game, maze_level_1b
    g = VGDLParser().parseGame(game_str)
    g.buildLevel(map_str)    
    senv = SubjectiveGame(g, actionDelay=100, recordingEnabled=True)
    #senv = GameEnvironment(g, actionDelay=100, recordingEnabled=True, visualize=True)
    task = GameTask(senv)    
    iagent = InteractiveAgent()
    exper = EpisodicExperiment(task, iagent)
    try:
        exper.doEpisodes(1)
    except UserTiredException:
        pass
    print senv._allEvents

def plotOptimalValues(gametype, layout, discountFactor=0.9, showValue=False):
    # build the game
    g = VGDLParser().parseGame(gametype)
    g.buildLevel(layout)
    
    # transform into an MDP
    C = MDPconverter(g)
    Ts, R, _ = C.convert()
    
    # find the optimal policy
    _, Topt = policyIteration(Ts, R, discountFactor=discountFactor)
    
    # evaluate the policy
    Vopt = trueValues(Topt, R, discountFactor=discountFactor)
        
    # plot those values    
    featurePlot((g.width, g.height), C.states, Vopt, plotdirections=True)
    
    if showValue:
        # expected discounted reward at initial state
        Vinit = Vopt[C.initIndex()]
        pylab.xlabel("V0=%.4f"%Vinit)